Gradient boosted decision trees reveal nuances of auditory discrimination behavior

Abstract

Perceptual constancy remains an elusive feature to measure in psychophysical animal studies due to the intractability of correlating trial parameters to trial outcomes and response times. Moreover, the nonlinear nature of our data constrains the potential that traditional statistical models have to elucidate the sensory and non-sensory factors that ultimately shape animal performance. In this study, we present an implementation of gradient-boosted regression and decision trees, a machine learning method, to gain insights into animal behavioural data. We implement this computational modelling method on a task in which 5 ferrets were trained to recognize a target word (“instruments”) against a stream of 54 other English non-target word tokens. To test perceptual constance to pitch, we manipulated the mean fundamental frequency (F0) of each trial across the whole series of word tokens (inter-trial F0 roving) or between word tokens (intra-trial F0 roving). We implemented Shapley Additive Explanations and permutation importance for model explainability to determine which factors were significant in our gradient-boosted regression and decision tree model performance. We found that, despite high performance on both intra-trial F0 roved and inter-trial F0 roved task paradigms, F0s that our subjects were introduced to after their initial control F0 period increased the rate of false alarms, and an F0 difference between the target word and the preceding non-target word caused slower reaction times. We also found a subset of non-target distractor words that increased the rate of false alarms and that shared spectrotemporal properties with the target word. Together, our results demonstrate that ferrets, like humans, are capable of showing tolerance to changes in pitch when discriminating complex sounds and that they rely on pitch as a cue to link sounds together over time. Finally, we illustrate the utility of gradient-boosted decision trees to gain insights into behavioural datasets.